IsiXhosa weather forecast generation

Building natural language generation systems for Nguni languages is difficult because traditional realisation solutions are incompatible and current extensions are unsuitable for general-purpose use. In this work, I will showcasing a general-purpose solution in the form of a novel model of grammar-infused templates. An evaluation on a weather forecast use case shows that the model is able to capture morphologically complex words that are outside the competence of basic/bare templates, generate texts that is completely free of morphological agreement and phonological conditioning errors, and a majority (57%) of its generated weather text is judged by L1 isiXhosa speakers as being fluent and grammaticality correct.